Artificial intelligence-based search method and apparatus, electronic device, and agent

Abstract

This disclosure provides a search method, apparatus, electronic device, and intelligent agent based on artificial intelligence, relating to the field of artificial intelligence technology, particularly to the fields of deep learning, large models, and big data technology. The method includes: acquiring search results matching text information input by an object, wherein the search results include information on multiple candidate dialogue intelligent agents, the candidate dialogue intelligent agent information including trigger entry information of the candidate dialogue intelligent agent, business information of the object to which the agent belongs, and response information generated by the candidate dialogue intelligent agent based on the business information in relation to the text information; processing the text information, the business information corresponding to the multiple candidate dialogue intelligent agents, and the response information using a large model to generate evaluation information for each of the multiple candidate dialogue intelligent agents; determining a target dialogue intelligent agent from the multiple candidate dialogue intelligent agents based on the multiple evaluation information; and displaying the trigger entry information of the target dialogue intelligent agent.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and more particularly to the fields of deep learning, large models and big data technology, specifically to an artificial intelligence-based search method, device, electronic device and intelligent agent. Background Technology

[0002] With the development of artificial intelligence technology, the application scenarios of AI based on Large Language Models (LLMs) are constantly increasing. For example, various information providers can use LLM-based agents to provide users with consulting or recommendation services.

[0003] However, given the vast amount of information available, it is crucial to utilize search engines to find suitable intelligent agents for users while meeting their search needs. Summary of the Invention

[0004] This disclosure provides a search method, apparatus, electronic device, and intelligent agent based on artificial intelligence.

[0005] According to one aspect of this disclosure, an artificial intelligence-based search method is provided, comprising: obtaining search results matching text information input by an object, wherein the search results include information on multiple candidate dialogue agents, the candidate dialogue agent information including trigger entry information of the candidate dialogue agent, business information of the object to which the candidate dialogue agent belongs, and response information generated by the candidate dialogue agent based on the business information in response to the text information; processing the text information, the business information and response information corresponding to the multiple candidate dialogue agents respectively using a large model to generate evaluation information for each of the multiple candidate dialogue agents; determining a target dialogue agent from the multiple candidate dialogue agents based on the multiple evaluation information; and displaying the trigger entry information of the target dialogue agent.

[0006] According to another aspect of this disclosure, an artificial intelligence-based search device is provided, comprising: an acquisition module for acquiring search results matching text information input by an object, wherein the search results include information on multiple candidate dialogue agents, the candidate dialogue agent information including trigger entry information of the candidate dialogue agent, business information of the object to which the candidate dialogue agent belongs, and response information generated by the candidate dialogue agent based on the business information in response to the text information; a generation module for processing the text information, the business information and response information corresponding to the multiple candidate dialogue agents respectively using a large model, and generating evaluation information for each of the multiple candidate dialogue agents; a determination module for determining a target dialogue agent from the multiple candidate dialogue agents based on the multiple evaluation information; and a display module for displaying the trigger entry information of the target dialogue agent.

[0007] According to another aspect of this disclosure, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method described above.

[0008] According to another aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are used to cause the computer to perform the method described above.

[0009] According to another aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements the method described above.

[0010] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0011] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein:

[0012] Figure 1 The illustration schematically depicts an exemplary system architecture applicable to artificial intelligence-based search methods and apparatus according to embodiments of the present disclosure.

[0013] Figure 2 A flowchart illustrating an artificial intelligence-based search method according to an embodiment of the present disclosure is shown.

[0014] Figure 3 An artificial intelligence-based search scenario according to an embodiment of the present disclosure is illustrated.

[0015] Figure 4A The diagram illustrates a data flow graph of the target dialogue agent selected according to a first embodiment of the present disclosure.

[0016] Figure 4B A data flow diagram illustrating the selection of a target dialogue agent according to a second embodiment of this disclosure is shown.

[0017] Figure 4C A data flow diagram illustrating the selection of a target dialogue agent according to a third embodiment of this disclosure is shown.

[0018] Figure 4D A data flow diagram illustrating the selection of a target dialogue agent according to a third embodiment of this disclosure is shown.

[0019] Figure 5The illustration depicts a scenario in which a search page is triggered according to an embodiment of the present disclosure.

[0020] Figure 6 The diagram illustrates a framework for generating evaluation rationale information and evaluation values ​​for a large model according to an embodiment of the present disclosure.

[0021] Figure 7 A block diagram of an artificial intelligence-based search device according to an embodiment of the present disclosure is shown schematically.

[0022] Figure 8 A schematic block diagram of an artificial intelligence agent according to an embodiment of the present disclosure is shown.

[0023] Figure 9 A block diagram of an electronic device suitable for implementing an artificial intelligence-based search method according to an embodiment of the present disclosure is illustrated schematically. Detailed Implementation

[0024] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0025] In the search field, search engines can calculate search relevance to provide results that are relevant to user input. Search relevance is a key technical indicator that characterizes the degree of content connection between a user's query and the search results, aiming to maximize the accuracy of the match between search results and the user's search intent. In some examples, the degree of matching between the user's input text and key information on the landing page can be used as search relevance. The more similar the key information on the landing page is to the user's input text, the more accurate the search result will be provided to the user from that landing page.

[0026] As a novel form of information delivery, intelligent agents typically provide services to users through dialogue. Landing pages are primarily in the form of LLM (Leadership Management System) chatbots, and their format, quality distribution, and features differ significantly from traditional landing pages. However, because landing pages are usually dialogue-based, they lack sufficient information for users to understand the services supported by the agent. Furthermore, it's difficult to obtain enough useful information from these landing pages for search relevance calculations, resulting in lower search accuracy for intelligent agents.

[0027] Furthermore, various information providers across different domains can offer business-related information through their respective agents. However, the general capabilities enabled by LLM allow agents belonging to multiple information providers to provide responses to user input text, unless the business is completely unrelated. In reality, these responses may not align with the user's intent, and the agent may struggle to provide relevant information through multiple rounds of dialogue. Users need to switch pages / agents to continue searching, resulting in a poor user experience.

[0028] To this end, embodiments of this disclosure provide an artificial intelligence-based search method to obtain search results that match text information input by an object. The search results include information on multiple candidate dialogue agents, which includes trigger entry information, business information of the object to which the candidate dialogue agent belongs, and response information generated by the candidate dialogue agent based on the business information. A large model is used to process the text information, the business information corresponding to each of the multiple candidate dialogue agents, and the response information to generate evaluation information for each candidate dialogue agent. Based on the multiple evaluation information, a target dialogue agent is determined from the multiple candidate dialogue agents. The trigger entry information of the target dialogue agent is then displayed.

[0029] In the above embodiments, for scenarios where services are provided through intelligent agents, the answer information of the intelligent agent and the business information of its affiliated object are obtained. Based on the above information and the recommendation reasons generated by the large model according to the above information and the text information input by the user, multiple candidate dialogue intelligent agents in the search results are filtered out as much as possible. Candidate dialogue intelligent agents that are irrelevant to the business and the text input by the user are eliminated, thereby improving search accuracy and user experience.

[0030] Figure 1 The illustration schematically depicts an exemplary system architecture applicable to artificial intelligence-based search methods and apparatus according to embodiments of the present disclosure.

[0031] It is important to note that Figure 1 The examples shown are merely examples of system architectures that can be applied to embodiments of this disclosure, intended to help those skilled in the art understand the technical content of this disclosure. However, they do not imply that embodiments of this disclosure cannot be used in other devices, systems, environments, or scenarios. For instance, in another embodiment, an exemplary system architecture for applying artificial intelligence-based search methods and apparatus may include a terminal device. However, the terminal device can implement the artificial intelligence-based search methods and apparatus provided by embodiments of this disclosure without interacting with a server.

[0032] like Figure 1As shown, the system architecture 100 according to this embodiment may include terminal devices 101, 102, and 103, a network 104, and a server 105. The network 104 serves as a medium for providing a communication link between the terminal devices 101, 102, and 103 and the server 105. The network 104 may include various connection types, such as wired and / or wireless communication links, etc.

[0033] Users can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, and 103, such as knowledge reading applications, web browser applications, search applications, instant messaging tools, email clients, and / or social platform software, etc. (for example only).

[0034] Terminal devices 101, 102, and 103 can be various electronic devices with displays and web browsing capabilities, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0035] Server 105 can be a server that provides various services, such as a backend management server that supports the content browsed by users using terminal devices 101, 102, and 103 (for example only). The backend management server can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0036] A server can be a cloud server, also known as a cloud computing server or cloud host, which is a host product in the cloud computing service system. It solves the shortcomings of traditional physical hosts and VPS services ("Virtual Private Server", or simply "VPS"), such as high management difficulty and weak business scalability. A server can also be a server for a distributed system or a server that incorporates blockchain technology.

[0037] It should be noted that the AI-based search method provided in this embodiment can generally be executed by terminal devices 101, 102, or 103. Accordingly, the large language model evaluation device provided in this embodiment can also be located in terminal devices 101, 102, or 103.

[0038] Alternatively, the AI-based search method provided in this embodiment can generally be executed by server 105. Correspondingly, the AI-based search device provided in this embodiment can generally be located in server 105. The AI-based search method provided in this embodiment can also be executed by a server or server cluster that is different from server 105 and capable of communicating with terminal devices 101, 102, 103 and / or server 105. Correspondingly, the AI-based search device provided in this embodiment can also be located in a server or server cluster that is different from server 105 and capable of communicating with terminal devices 101, 102, 103 and / or server 105. Alternatively, the AI-based search method provided in this embodiment can generally be executed by terminal devices 101, 102, or 103. Correspondingly, the AI-based search device provided in this embodiment can generally be located in terminal devices 101, 102, or 103.

[0039] For example, a user can obtain text information input by interacting with terminal devices 101, 102, and 103. Terminal devices 101, 102, and 103 send the obtained text information to a search engine to retrieve search results. The search engine can be located within server 105. Server 105 retrieves the search results and calls a large model to generate evaluation information. Based on the evaluation information, it determines the target dialogue agent and its trigger entry information. Then, server 105 can send the target dialogue agent's trigger entry information to terminal devices 101, 102, and 103 for display.

[0040] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0041] In the embodiments disclosed herein, the collection, storage, use, processing, transmission, provision, disclosure, and application of any type of information, such as user personal information, comply with the provisions of relevant laws and regulations, necessary confidentiality measures have been taken, and they do not violate public order and good morals.

[0042] In the technical solution disclosed herein, the user's authorization or consent is obtained before acquiring or collecting the user's personal information.

[0043] Figure 2 A flowchart illustrating an artificial intelligence-based search method according to an embodiment of this disclosure is shown schematically. Figure 2 As shown, the embodiment includes operations S210 to S240.

[0044] In operation S210, search results matching the text information input by the object are obtained. The search results include information on multiple candidate dialogue agents. The information on candidate dialogue agents includes the trigger entry information of the candidate dialogue agent, the business information of the object to which the agent belongs, and the response information of the candidate dialogue agent to the text information generated by the candidate dialogue agent based on the business information.

[0045] In operation S220, a large model is used to process text information, business information and response information corresponding to multiple candidate dialogue agents, and to generate evaluation information for each of the multiple candidate dialogue agents.

[0046] In operation S230, the target dialogue agent is determined from multiple candidate dialogue agents based on multiple evaluation information.

[0047] When operating S240, the trigger entry information of the target dialogue agent is displayed.

[0048] Text information represents a user's (e.g., an object's) search intent. Users can input text information through the search engine's interactive interface. For example, text information could be "the effects of herb A," "what does XX mean," or "how can I buy item B," etc.

[0049] Search results can be determined by a search engine based on text information entered by the user. Search engines can support searching for candidate dialogue agents as well as searching traditional landing pages. Search engines can determine search results that match the input text information in various ways. For example, if information exists within the landing page, keyword matching can be performed between the text information and the information on the landing page; alternatively, models such as LLM can be used for searching, and the search results here can serve as a simple initial screening result.

[0050] Candidate dialogue agents are artificial intelligence entities capable of interacting with users in natural language to understand user intent and provide targeted services or information. Candidate dialogue agents typically possess core capabilities such as dialogue understanding, intent recognition, knowledge integration, and response generation.

[0051] Candidate dialogue agent information represents the information corresponding to the candidate dialogue agent. Trigger entry information is used to trigger the searched candidate dialogue agent; this trigger entry information can be a link, button, icon, or other interactive element. Users can enter the landing page of the candidate dialogue agent, such as a dialogue window, through clicking, touching, or other preset operations. In addition, candidate dialogue agent information may also include the agent's identification information, domain tags, brief description, user rating stars, etc.

[0052] The attributing entity can be an information provider. Multiple attributing entities can provide multiple candidate dialogue agents to offer users responses that align with their business needs.

[0053] In some embodiments, the attributing entity can fine-tune a basic or general candidate dialogue agent using its business information or training data constructed based on business information, so that the fine-tuned candidate dialogue agent can generate response information that is business-oriented in response to the same input text information. For example, if the attributing entity is used to provide literature services, it can fine-tune the candidate dialogue agent using information related to the literature. Then, if the input text information is "the effect of herb A", the response information generated by the candidate dialogue agent could be "Regarding the role of herb A in the XX field, a new literature has recently emerged, which explains the role of herb A. The new literature is YYY".

[0054] The business information of the target entity describes the business it provides. For example, business information may include product information, service content, professional knowledge, business processes, promotional activities, and other information related to the target entity's business. For instance, the business information of a target entity providing educational resources could be: "Main Business: XXX is an online education platform providing XXX. Its business covers multiple areas such as A, B, C, and D. It primarily targets the XXX group. Its service advantages include XXXXX."

[0055] In one embodiment, the aforementioned attribution object may be an advertiser, and the intelligent agent of the attribution object may be an advertising intelligent agent used to provide services within the advertiser's business scope.

[0056] Large models can be pre-trained, high-parameter models with massive amounts of knowledge and powerful natural language understanding and generation capabilities. These large models, trained on large-scale text corpora, can learn the syntax, semantics, contextual relationships, and world knowledge of a language. In the embodiments of this disclosure, text information, business information, and response information can be used as inputs to the large model. Leveraging the large model's language understanding capabilities, and combining the business information of the belonging object with the response information of the candidate dialogue agent, it is determined whether the candidate dialogue agent matches the user's input text information, and evaluation information is generated.

[0057] Evaluation information can be information used to characterize whether a candidate dialogue agent matches the text information input by the user and / or information characterizing the degree of matching between the candidate dialogue agent and the text information input by the user. For example, evaluation information can be numerical information to reflect whether the candidate dialogue agent matches the text information and the degree of matching through quantifiable numerical values; or, evaluation information can be textual information to reflect whether the candidate dialogue agent matches the text information and / or the degree of matching through interpretable text.

[0058] The target dialogue agent can be a candidate dialogue agent that matches the text information input by the user from among multiple candidate dialogue agents. There can be one or more target dialogue agents, and the number of target dialogue agents can be determined according to a predetermined search rule. If there are multiple target dialogue agents, the trigger entry information of each target dialogue agent can be displayed in order of the degree of matching between the candidate dialogue agents represented by the evaluation information and the input information. For example, multiple trigger entry information can be displayed in descending order of matching degree.

[0059] In some embodiments, trigger entry information for at least one target dialogue agent can be displayed in the search engine. Users can browse at least one target dialogue agent in the search engine and click on the trigger entry information of a target dialogue agent to enter the landing page of that target dialogue agent for dialogue.

[0060] In the embodiments of this disclosure, for the search scenario of candidate dialogue agents, the business information of the candidate dialogue agent's affiliated object can compensate for the lack of information of the candidate dialogue agent within the dialogue page, thereby providing more complete and business-related information for subsequent candidate dialogue agent evaluation. Then, using a large model, evaluation information for each candidate dialogue agent is determined based on the business information, the candidate dialogue agent's response information, and the user's input text information. This improves the evaluation accuracy of candidate dialogue agents in the search results from multiple perspectives: the text information input by the user, the response information output by the candidate dialogue agent, and the business information upon which the output depends. Subsequently, when filtering and displaying target dialogue agents using the evaluation information, candidate dialogue agents that are irrelevant to the user's input text information can be filtered out as much as possible. This reduces the number and time users spend browsing irrelevant candidate dialogue agents during the search process, thereby improving search accuracy and user experience, and helping users quickly find candidate dialogue agents that meet their needs.

[0061] Figure 3 The illustration schematically depicts an artificial intelligence-based search scenario according to embodiments of the present disclosure. For example... Figure 3As shown, the search engine can provide services through interactive page 310. Users can input text information through interactive controls 311 on interactive page 310. These interactive controls 311 can be various controls such as input boxes and voice input buttons. After the user inputs the text information "the effects of herb A", the search engine can obtain search results that match the text information. If the search results include 10 candidate dialogue agents, after determining the evaluation information for the 10 candidate dialogue agents through the method provided in this embodiment, a target dialogue agent can be determined from the 10 candidate dialogue agents. The business information of the target dialogue agent can be "providing literature search services in the field of pharmaceuticals".

[0062] Therefore, the trigger entry information 313 of the target dialogue agent can be displayed on the interactive page 310. In this embodiment, the trigger entry information 313 may include the response information to the text information, such as "A new literature has recently emerged explaining the role of herb A in the XX field, and the new literature is YYY". Clicking the trigger entry information 313 will allow the user to enter the dialogue window 320 of the target dialogue agent. Within this dialogue window, text information 321 and response information 323 can be automatically displayed. Furthermore, within this dialogue window, the user can continue to input text information through the dialogue control 324 to achieve multi-round interactive question-and-answer under the intent of literature consultation.

[0063] In addition, the search results may include trigger entry information 312 of the traditional landing page, and users can enter the introduction page of herb A by clicking on the trigger entry information 312.

[0064] According to an embodiment of this disclosure, for operation S220, a large model is used to process text information, business information corresponding to multiple candidate dialogue agents, and answer information to generate evaluation information for each of the multiple candidate dialogue agents. This includes: selecting at least one first candidate dialogue agent from the multiple candidate dialogue agents based on the semantic similarity between the text information and the business information; and for each first candidate dialogue agent, a large model is used to process text information, answer information corresponding to the first candidate dialogue agent, and business information to generate evaluation information for the first candidate dialogue agent.

[0065] Textual information and business information can be converted into text vectors respectively, and the similarity between two text vectors can be used as semantic similarity. For example, candidate dialogue agents with semantic similarity greater than a certain semantic threshold, such as 0.8, can be selected as the first candidate dialogue agent; conversely, candidate dialogue agents with semantic similarity less than a certain semantic threshold can be selected as the second candidate dialogue agent.

[0066] For example, taking the input text "Medicinal components and clinical applications of herb A" as an example, the search results include three candidate dialogue agents 1-3, whose respective business information is "Research on drug growth environment", "Research on drug medicinal value", and "Drug encyclopedia answers". Due to the general language understanding and response capabilities of the agents themselves, all three candidate dialogue agents can generate corresponding answers to the text information. For example, they could be "Herb A includes XXX components, which are usually grown in a CCC environment", "Herb A includes XXX components, and its clinical applications are XYYY", or "Herb A includes XXX components, component 1, component 2...". While the answers provided by these three candidate dialogue agents superficially address the user's questions to some extent, the user's actual intent lies primarily in "medicinal components and clinical applications." Candidate dialogue agent 2 can provide further information to the user. Figure 1 The subsequent response was then provided. Therefore, semantic similarity was first calculated based on the business information of the target entity to which each candidate dialogue agent belonged and the input text information, thus selecting two first candidate dialogue agents, such as candidate dialogue agents 1 and 2.

[0067] For the first candidate dialogue agents obtained after screening based on business information, the text information, response information, and business information are then used as inputs to the large model, which generates evaluation information for each first candidate dialogue agent.

[0068] In the above embodiments, the input to the large model includes user-side information (text information) and agent-side information (response information and business information). For scenarios where the landing page of the candidate dialogue agent lacks key information, the large model can comprehensively evaluate whether the information (business information and response information) corresponding to the candidate dialogue agent closely matches the user's text information, based on its deep understanding of natural language and its cognitive ability regarding business scenarios. For example, it can evaluate whether the business information and response information semantically match the user's actual needs, thus obtaining evaluation information.

[0069] In the embodiments of this disclosure, through a hierarchical evaluation design, the range of candidate dialogue agents is first narrowed down from the business dimension by rapid semantic similarity calculation, reducing the resource consumption of subsequent large-scale model evaluation; at the same time, the large-scale model is used to conduct a fine evaluation of the deep semantic relationship between business information, answer information and user text information, so that the first candidate dialogue agent selected is not only related to the user's needs in terms of surface semantics, but also closely matches the user's intent in terms of business logic and actual answer results, thereby improving the search accuracy of candidate dialogue agents.

[0070] According to embodiments of this disclosure, the evaluation information includes: evaluation value and evaluation reason information; using a large model to process text information, response information corresponding to the first candidate dialogue agent, and business information to generate evaluation information for the first candidate dialogue agent, including: combining the response information corresponding to the first candidate dialogue agent, business information, and text information corresponding to the object into prompt information and inputting it into the large model, so as to use the large model to perform a discrimination task and a generation task respectively to obtain the evaluation value and evaluation reason information.

[0071] The response information, business information, and text information corresponding to the first candidate dialogue agent are combined to form a prompt, which serves as the input to the large model.

[0072] In some embodiments, the prompt information may have a predetermined template or combination rules. For example, to facilitate the large model in distinguishing between agent-side and user-side information, the response information, business information, and text information can be arranged in the order of user-side first, then agent-side. Alternatively, the agent-side and user-side information can be separated by a delimiter. The agent-side information includes at least the response information and business information corresponding to the first candidate dialogue agent, which form the agent-side feature domain; the user-side information includes at least text information.

[0073] The evaluation value is used to quantify the degree of matching between the information corresponding to the first candidate dialogue agent and the information corresponding to the object. The discrimination task can be a classification task, classifying the degree of matching between the information corresponding to the first candidate dialogue agent and the information corresponding to the object. In one embodiment, the discrimination task can classify the degree of matching between information combining answer information and business information and the user needs represented by the text information. It is understood that the discrimination task can be a multi-classification task; after determining the classification result based on the current input, the evaluation value is output based on the correspondence between the classification result and a predetermined evaluation value. For example, evaluation values ​​1, 2, 3, 4, and 5 represent five categories: no match, low match, moderate match, high match, and extremely high match, respectively.

[0074] Evaluation justification information is used to explain the evaluation value. The task can generate evaluation justification information based on business information, response information, and text information, assessing the degree of match between the information corresponding to the first candidate dialogue agent and the user's needs represented by the information (such as text information) corresponding to the object. For example, if an evaluation value of 4 indicates that the response information of the first candidate dialogue agent meets the user's needs, the evaluation justification information could be: specifically explaining which content in the response information, and from which angles, accurately responds to the user's needs represented by the text information.

[0075] In some embodiments, large models can perform discrimination and generation tasks in parallel to improve evaluation efficiency.

[0076] In another embodiment, the large model can also perform the discrimination and generation tasks sequentially. For example, the generation task can be performed first, and the evaluation reasoning information obtained from the generation task can be used as supplementary information for the semantic dimension. This information, along with textual information, response information, and business information, can be used as input for the discrimination task to improve the accuracy of the evaluation values, thereby improving the search accuracy of the subsequent target dialogue agent.

[0077] In the embodiments of this disclosure, a large model is used to perform discrimination and generation tasks based on business information, response information, and text information, respectively, and to obtain evaluation values ​​and evaluation reasoning information. The evaluation information can be output from quantifiable and interpretable dimensions, so that multiple first candidate dialogue agents can be screened or ranked based on quantifiable evaluation values; and the large model can be optimized based on interpretable evaluation reasoning information.

[0078] In some embodiments, in addition to displaying the trigger entry information of the target dialogue agent, evaluation values ​​and evaluation rationale information may also be displayed so that users can intuitively see and understand the selection criteria for the target dialogue agent.

[0079] According to a specific embodiment of this disclosure, determining a target dialogue agent from multiple candidate dialogue agents based on multiple evaluation information includes: selecting a target dialogue agent from at least one first candidate dialogue agent based on a comparison relationship between an evaluation value and an evaluation threshold.

[0080] The evaluation threshold can be predetermined, such as 2 or 3, to filter target dialogue agents based on the comparison between the evaluation value and the evaluation threshold. For example, the first candidate dialogue agent whose evaluation value is greater than the evaluation threshold can be selected as the target dialogue agent.

[0081] In other embodiments, at least one first candidate dialogue agent can be sorted according to the evaluation value to select at least one first candidate dialogue agent with the largest evaluation value from at least one first candidate dialogue agent and use it as the target dialogue agent.

[0082] In the embodiments of this disclosure, by using the comparison relationship between quantifiable evaluation values ​​and predetermined thresholds, multiple first candidate dialogue agents can be screened or sorted in order to provide the user with trigger entry information of a target dialogue agent that better meets the user's needs.

[0083] Figure 4A A data flow diagram illustrating the selection of the target dialogue agent according to a first embodiment of this disclosure is shown schematically. Figure 4AAs shown, the search engine can obtain search results based on the text information 410 input by the user. These search results can include information on multiple candidate dialogue agents. Based on the identification information contained in the candidate dialogue agent information, multiple candidate dialogue agents can be identified, such as candidate dialogue agent 1 M-1... candidate dialogue agent 1 MN, for a total of N candidate dialogue agents. Each of the N candidate dialogue agents is associated with business information of its respective affiliated object. Inputting the text information 410 into each of the N candidate dialogue agents will yield N response messages. N is a positive integer greater than or equal to 2.

[0084] Taking candidate dialogue agent 1 M-1 as an example, candidate dialogue agent 1 M-1 is associated with business information 421 of the belonging object. Inputting text information 410 into candidate dialogue agent 1 M-1 yields response information 431. Using the large model M110, based on text information 410, business information 421 of the belonging object, and response information 431, evaluation information 441 for candidate dialogue agent 1 M-1 can be generated. Evaluation information 441 includes evaluation value and evaluation reason information. As mentioned above, the large model M110 can be used to calculate the semantic similarity between text information 410 and business information 421 to select the first candidate dialogue agent. If candidate dialogue agent 1 M-1 is the first candidate dialogue agent, the large model M110 can be used to perform discrimination and generation tasks respectively to obtain evaluation value and evaluation reason information.

[0085] For each first candidate dialogue agent, corresponding evaluation information can be obtained. Based on this evaluation information, the target dialogue agent can be determined from multiple candidate dialogue agents. Taking candidate dialogue agent 1M-1 as an example, if the evaluation value in the evaluation information of candidate dialogue agent 1M-1 is greater than a predetermined threshold, then candidate dialogue agent 1M-1 can be determined as the target dialogue agent, and the trigger entry information of the target dialogue agent can be displayed so that the user can interact with it.

[0086] According to embodiments of this disclosure, the method further includes: searching a knowledge base for domain background information and / or word explanation information matching the text information, wherein the domain background information is determined based on semantic search of the text information, and the word explanation information is determined based on semantic search of at least one keyword in the text information; taking at least one of the domain background information and word explanation information corresponding to the object and the text information as information corresponding to the object, and combining the information corresponding to the object, the answer information corresponding to the first candidate dialogue agent, and the business information into prompt information; inputting the prompt information into a large model to perform a discrimination task and a generation task using the large model respectively, and obtaining an evaluation value and evaluation reason information.

[0087] To improve the large model's ability to understand user needs expressed in textual information, further information corresponding to the user can be added, such as domain background information and / or word explanation information.

[0088] Domain background information is used to represent information related to the professional field to which the text information belongs, such as industry knowledge, technical background, or application scenarios, so that the large model can understand the user's input text information based on the specific context of the field. Word explanation information is used to explain the keywords appearing in the text information, such as definition information and / or example information of keywords, so that the large model can accurately grasp the specific meaning and semantic connotation of these words in the current text information.

[0089] The knowledge base can pre-store structured information from multiple domains. For user-inputted text information, structured information with semantic similarity to the text information can be searched from the knowledge base at the sentence level and used as domain background information; alternatively, the text information can be segmented and keywords extracted, and for one or more keywords extracted from the text information, structured information with semantic similarity to the keywords can be searched from the knowledge base at the word level and used as word explanation information.

[0090] In some embodiments, domain background information and / or word explanation information that match the text information can be searched from the knowledge base based on Retrieval Augmented Generation (RAG).

[0091] When domain background information and / or word explanation information are available, this information can also be used as the information corresponding to the object, also known as user-side information. Similar to the operation of the large model generating evaluation information based on text information, business information, and response information, at least one of the domain background information and word explanation information, along with the text information, can be used as the information corresponding to the object. The large model can then generate evaluation information based on the information corresponding to the object, as well as the business information and response information corresponding to the first candidate dialogue agent.

[0092] In the embodiments of this disclosure, by adding domain background information and / or word explanation information that matches the text information to the large model input, richer and more accurate contextual support can be provided for the understanding and reasoning process of the large model, avoiding the impact of lack of domain common sense or ambiguity on the evaluation accuracy, thereby improving the search accuracy of the subsequent target dialogue agent and the user experience.

[0093] Figure 4B A data flow diagram illustrating the selection of a target dialogue agent according to a second embodiment of this disclosure is shown. Figure 4BThe system can search the knowledge base for domain background information 451 and word explanation information 452 that match the text information 410, and use the domain background information 451, word explanation information 452, and text information 410 as information corresponding to the object, and the business information 421 of the object and the answer information as information corresponding to the candidate dialogue agent, and input them together into the large model M110. Figure 4B The operations for acquiring information other than domain background information 451 and word explanation information 452, as well as the operations for generating evaluation information from the large model and filtering to obtain the target dialogue agent, can be found in [link to relevant documentation]. Figure 4A The description will not be repeated here.

[0094] According to embodiments of this disclosure, the prompt information includes a first separator and at least one second separator. The first separator is used to separate information corresponding to the object and information corresponding to the first candidate dialogue agent. The second separator is used to separate business information and answer information. The second separator is also used to separate at least two of text information, domain background information, and word explanation information. At least one piece of information corresponding to the object and at least one piece of information corresponding to the first candidate dialogue agent in the prompt information are arranged in a predetermined order.

[0095] For example, the prompt message could be: "ABB Training Fees | Education and Language Training | ABB Course Fee Analysis [SEP] Education and K-12 Education | Main Business: XXX is an online education platform providing XXX. Its business covers multiple directions such as A, B, C, and D. It mainly targets the XXX group. Service advantages include XXXXX. | ABB training fees vary depending on the subject, generally ranging from xxxx. Please refer to the actual situation of the school for specific fees." "ABB Training Fees," "Education and Language Training," "ABB Course Fee Analysis," "Education and K-12 Education," "Main Business..." and "ABB Training Fees... Subject to actual situation of the school" represent text information, domain background information, term explanation information, description information of the target audience, business description information of the target audience, and response information, respectively. The business information of the target audience can include the description information of the target audience and / or the business description information of the target audience. "[SEP]" can be the first separator, and "|" can be the second separator. The information before "[SEP]" corresponds to the target audience, and multiple pieces of information corresponding to the target audience are separated by the second separator "|". The information following “[SEP]” corresponds to the first candidate dialogue agent, and the multiple pieces of information corresponding to it are also separated by the second separator “|”.

[0096] The predetermined order of multiple pieces of information in the prompt message is as follows: information corresponding to the object is placed before information corresponding to the first candidate dialogue agent; text information, domain background information, and word explanation information are arranged in order; business information of the object to which the prompt belongs and response information of the first candidate dialogue agent are arranged in order.

[0097] In the embodiments of this disclosure, by utilizing the first separator, the second separator, and a predetermined order of multiple pieces of information, the information corresponding to the object and the information corresponding to the first candidate dialogue agent can be structurally distinguished in the input of the large model. This improves the logic and specificity of the large model in the semantic understanding process, enabling the large model to quickly identify and distinguish multiple pieces of information from the prompt information. It avoids the semantic mixing in the prompt information from affecting the large model's specific analysis of each piece of information, thereby improving the accuracy and efficiency of the large model's output evaluation information.

[0098] According to embodiments of this disclosure, before using a large model to process text information, response information corresponding to a first candidate dialogue agent, and business information to generate evaluation information for the first candidate dialogue agent, the method further includes: using a large model to simplify the response information based on the business information to obtain simplified response information, wherein the simplified response information, together with the business information and the response information, is used to generate evaluation information.

[0099] The input to the large model includes information from two dimensions: the user and the candidate dialogue agents. Some dimensions also include multiple pieces of information, such as the first candidate dialogue agent including corresponding business information and response information. However, as the input information to the large model gradually increases, important information may be overwhelmed by other information, leading to biases in the evaluation information output by the large model.

[0100] For example, a candidate dialogue agent can generate response information based on textual information, but the business orientation of this response information depends on the business information of the target audience. If there is too much input information into the large model, other information will mask or weaken the core business information, leading to bias in the evaluation information output by the large model.

[0101] In the embodiments of this disclosure, the input to the large model includes: business information, text information input by the user, and response information generated by the candidate dialogue agent. The text information input by the user is usually relatively short, and the length of the response information generated by the candidate dialogue agent is related to the setting of the belonging object, and its length is usually not fixed. Therefore, before generating evaluation information, the large model simplifies the response information generated by the candidate dialogue agent based on the business information. Then, the simplified response information can replace the above text information, so that the large model generates evaluation information based on the text information, business information, and simplified response information.

[0102] In one embodiment, the business information can be used as a semantic benchmark to extract and compress the answer information at the semantic level; or, text information with low semantic similarity to the business information in the answer information can be removed to obtain simplified business information.

[0103] In another embodiment, the text length of the business information can be used as a length benchmark. Simplified business information is generated based on the response information, and the text length of the simplified business information falls within a length range that matches the length benchmark. For example, the boundaries of this length range could be 0 or one-third of the text length of the business information.

[0104] In the embodiments of this disclosure, the large model is used to simplify the answer information. This not only avoids the business information from being obscured by excessively long text information in the input of the large model, making it easier for the large model to capture core business features, but also retains important information in the answer information, so as to improve the accuracy of subsequent evaluation information generation and thus improve search accuracy.

[0105] According to a specific embodiment of this disclosure, a large model is used to simplify the response information based on business information to obtain simplified response information. This includes: determining dynamic text constraints that match the text length of the business information, wherein the dynamic text constraints are used to limit the text length of the response information; and using the large model to simplify the response information based on the dynamic text constraints to obtain simplified response information.

[0106] Dynamic text constraints can be understood as a simplified text length benchmark for the response information. Multiple text length matching rules can be pre-defined. Upon receiving business information, the text length of the business information is first calculated; for example, the number of characters, words, or sentences contained in the business information can be counted. Then, based on this text length, the corresponding dynamic text constraint is selected from the pre-defined text length matching rules.

[0107] For example, the text length matching rule could be: if the text length is between 50 and 150 characters, the simplified response text length should not exceed 50% of the business information text length; if the text length is between 200 and 300 characters, the simplified response text length should not exceed 40% of the business information text length. Therefore, if the business information text length is 100 characters, the dynamic text constraint can be determined as: the simplified text length limit is 50 characters; if the business information text length is 200 characters, the dynamic text constraint is: the simplified text length limit is 80 characters.

[0108] Furthermore, to avoid oversimplification leading to the loss of key information in the response, the text length matching rule can also include: the simplified response text length is greater than 10% of the business information text length. In this case, if the business information text length is 100 characters, the dynamic text constraint is: the simplified text length upper limit is 50 characters, and the text length lower limit is 10 characters.

[0109] It is understandable that the dynamic text constraints correspond to the specific business information in each input, and when evaluating multiple candidate dialogue agents, the text length of the input to the large model is different.

[0110] Large models can simplify response information by performing summarization tasks. In some embodiments, after determining dynamic text constraints, these constraints can be used as one of the inputs for the large model to perform the summarization task, guiding the large model to simplify business information under these constraints.

[0111] For example, instead of the answer "ABB training fees vary depending on the subject, generally ranging from xxxx. Please refer to the actual situation of the campus for specific fees," a simplified answer could be "ABB training fees are generally in the range of xxxx, and the specific fees are related to the subject and campus."

[0112] In the embodiments of this disclosure, by using dynamic text constraints that match the length of the business information text, the simplification of the response information can be controlled more flexibly and specifically. This can effectively shorten the text length and adapt to business information of different lengths. It can not only avoid the business information being obscured by excessively long text information in the input of large models, but also take into account the richness of the input of large models, thereby improving the accuracy of the generated evaluation information.

[0113] Figure 4C A data flow diagram illustrating the selection of the target dialogue agent according to a third embodiment of this disclosure is shown schematically. Figure 4C As shown, for candidate dialogue agent 1 M-1, the input of the large model M110 includes: text information 410, business information of the belonging object 421, and response information 431.

[0114] The large model M110 can first perform a summary generation task based on the business information 421 of the belonging object and the response information 431 to simplify the response information 431 and obtain the simplified response information. Then, it can use the text information 410, the business information 421 of the belonging object, and the simplified response information to perform a discrimination task and a generation task respectively to obtain the corresponding evaluation value and evaluation reason information.

[0115] Figure 4CFor operations involving the acquisition of various types of information, the generation of evaluation information, and the selection of target dialogue agents, please refer to [link to relevant documentation]. Figure 4A The description will not be repeated here.

[0116] According to embodiments of this disclosure, the process of using a large model to process text information, response information corresponding to a first candidate dialogue agent, and business information to generate evaluation information for the first candidate dialogue agent includes: using the large model to perform business conflict detection on the business information to obtain a business conflict detection result; and when the business conflict detection result indicates that there are at least two conflicting business sub-information in the business information, using the large model to process the text information, response information corresponding to the first candidate dialogue agent, and at least two business sub-information to generate at least two evaluation information for the first candidate dialogue agent, wherein the evaluation information corresponds to the business sub-information.

[0117] Business conflict detection is used to detect whether there are conflicting business scopes in business information. For example, conflicting business scopes may refer to group conflicts among multiple businesses belonging to the same entity, opposite business experiences, or significant differences in business content.

[0118] For example, a business entity providing educational services could offer tutoring for primary and secondary school students and academic improvement services for adults, or free courses and paid courses. "Providing tutoring for primary and secondary school students and academic improvement services for adults" could be seen as two services addressing group conflicts, while "providing free courses and paid courses" could be seen as two services with opposite user experiences. Furthermore, an online platform entity could offer video browsing and shopping services; these two services could be considered types with significantly different content.

[0119] Business conflict detection can be achieved using large-scale models to understand the semantics of business information. For example, large-scale models can perform deep semantic parsing of business sub-information within business information, identifying the business domain, target audience, business experience, and business content involved in multiple business sub-information, and then performing semantic comparisons between multiple business sub-information across these dimensions to obtain business conflict detection results.

[0120] The business conflict detection results can include whether there are conflicting business sub-information within the business information, and which business sub-information is conflicting. It is understood that there can be two or more conflicting business sub-information pieces within the business information. Regardless of the number of conflicting business sub-information pieces, subsequent business information splitting can be performed when conflicting business sub-information exists.

[0121] For example, business information can be broken down into at least two conflicting business sub-information based on business conflict detection results. This operation treats a real first-candidate dialogue agent as at least two virtual third-candidate dialogue agents, where the number of third-candidate dialogue agents is the same as the number of conflicting business sub-information. Then, for each third-candidate dialogue agent, the business sub-information is treated as the only business information supported by that third-candidate dialogue agent, and evaluation information is generated for each third-candidate dialogue agent using a method similar to that used to generate evaluation information. It is understood that the evaluation information of at least two third-candidate dialogue agents is essentially the evaluation information of the first-candidate dialogue agent within the conflicting business scope.

[0122] In the embodiments of this disclosure, by assigning conflicting business sub-information to different virtual third candidate dialogue agents and generating evaluation information for each third candidate dialogue agent, the originally potentially ambiguous overall evaluation can be broken down into specific evaluations for different conflicting branches, making the evaluation information more targeted and interpretable, thereby improving the search accuracy of subsequent candidate dialogue agents.

[0123] According to embodiments of this disclosure, determining and displaying a target dialogue agent from multiple candidate dialogue agents based on multiple evaluation information includes: for a first candidate dialogue agent having at least two evaluation information, if there is evaluation information in the at least two evaluation information whose evaluation value is greater than an evaluation threshold, determining the first candidate dialogue agent as the target dialogue agent.

[0124] For a first candidate dialogue agent with at least two evaluation information, if the evaluation value in one of the evaluation information is greater than the evaluation threshold, it indicates that the business sub-information corresponding to the evaluation information is more in line with the user's needs. The first candidate dialogue agent has stronger adaptability in processing the business sub-information, and the first candidate dialogue agent can be identified as the target dialogue agent and displayed.

[0125] If there is no evaluation information with an evaluation value greater than the evaluation threshold among at least two evaluation information, then the first candidate dialogue agent will not be determined as the target dialogue agent.

[0126] In this embodiment, by splitting conflicting business sub-information and generating targeted evaluation information, the accuracy of searching for the target dialogue agent is improved while avoiding ambiguity in business requirements caused by conflicting business sub-information.

[0127] Figure 4D A data flow diagram illustrating the selection of the target dialogue agent according to a third embodiment of this disclosure is shown schematically. Figure 4DAs shown, taking candidate dialogue agent 1 M-1 as an example, business conflict detection can be performed on the business information 421 of the belonging object. If the business conflict detection result indicates that there are P conflicting business sub-information in the business information, the business information 421 of the belonging object can be split into P business sub-information, such as business sub-information 1 460-1... business sub-information P 460-P, where P can be a positive integer greater than or equal to 2.

[0128] Taking business sub-information P 460-P as an example, the response information 431 and business sub-information P 460-P can be regarded as information corresponding to the virtual candidate dialogue agent 1-P M-1-P. This information, along with text information 410, is input into the large model M110 to obtain evaluation information 441-P for the candidate dialogue agent 1-P. Similarly, other business sub-information can also obtain evaluation information in the same way. If the evaluation value in evaluation information 441-P is greater than the evaluation threshold, then the candidate dialogue agent 1 M-1 corresponding to the virtual candidate dialogue agent 1-P M-1-P can be determined as the target dialogue agent. Figure 4D For operations involving the acquisition of various types of information, the generation of evaluation information, and the selection of target dialogue agents, please refer to [link to relevant documentation]. Figure 4A The description will not be repeated here.

[0129] According to embodiments of this disclosure, the method further includes: when there is a second agent other than the first candidate dialogue agent among a plurality of candidate dialogue agents, using a predetermined evaluation value and predetermined text information as the evaluation value and evaluation reason information of the second agent, respectively, wherein the predetermined evaluation value is less than or equal to the evaluation value generated using a large model.

[0130] After calculating the semantic similarity between the business information and text information of the candidate dialogue agents' respective objects, candidate dialogue agents with semantic similarity less than a certain semantic threshold can be designated as second candidate dialogue agents. For the second candidate dialogue agents, instead of using a large model to generate corresponding evaluation information, the predetermined evaluation information can be directly used as the evaluation information for the second agent.

[0131] For example, if a large model can generate an evaluation value of 1 to 5 when performing a discrimination task, then the predetermined evaluation value can be 0, and the predetermined text information can be: The information provider of the current candidate dialogue agent does not support the business of the text information input by the user.

[0132] In the embodiments of this disclosure, by directly using predetermined evaluation values ​​and predetermined text as the evaluation value and evaluation reason information of the second agent, respectively, quantifiable and interpretable evaluation information can be provided to the second candidate dialogue agent through predetermined evaluation values ​​and predetermined text information without using a large model to generate evaluation information.

[0133] According to embodiments of this disclosure, the candidate dialogue agent information also includes the identification information of the candidate dialogue agent. The process of obtaining search results that match the text information input by the object includes: obtaining the business information of the home object of each candidate dialogue agent from the home object maintenance library based on the identification information and trigger entry information of each candidate dialogue agent.

[0134] Search engines can determine search results that match the input text information in various ways. The matching results can include the identification information and trigger entry information of each of the multiple candidate dialogue agents that match the text information.

[0135] The attribution object maintenance library can be a database used to maintain information related to attribution objects. For example, there is a correspondence between the identification information of candidate dialogue agents and their attribution objects, and there is also a correspondence between attribution objects and business information. Both of these correspondences can be stored in the attribution object maintenance library.

[0136] For example, the identification information of candidate dialogue agents can be used to determine business information based on the two correspondences mentioned above. Alternatively, the ownership object maintenance library can maintain the information corresponding to the ownership object as a chain-like correspondence, and then query the ownership object's business information based on the identification information of the candidate dialogue agents and the chain-like correspondence.

[0137] Considering that the tendency of candidate dialogue agents to generate answers has been fixed within the model parameters of the candidate dialogue agents using training data, and that the landing pages of the dialogue formats adopted by the candidate dialogue agents lack clear business information, the embodiments of this disclosure, through the maintenance library of the belonging objects, can quickly obtain the business information of the belonging objects of each candidate dialogue agent in the search results. This not only compensates for the lack of information on the landing pages, but also uses the searched business information to generate more accurate evaluation information, and uses the more accurate evaluation information to filter agents, thereby improving search accuracy.

[0138] According to embodiments of this disclosure, obtaining search results that match the text information input by the object further includes: obtaining historical dialogue information of each candidate dialogue agent from the object maintenance library based on the identification information of each candidate dialogue agent; determining historical text information similar to the text information from the historical dialogue information of each candidate dialogue agent; and using historical response information for the historical text information as response information for the text information.

[0139] The ownership object maintenance library can also include historical dialogue information of candidate dialogue agents. For example, the ownership object maintenance library can store historical dialogue information of candidate dialogue agents through a separate sub-library.

[0140] Historical dialogue information includes the historical text information of the candidate dialogue agents and the historical response information generated based on that historical text information. Similar to the text information mentioned above, historical text information can be text information that the user entered through the search engine before the current time period, such as "How can herb A be used?"

[0141] For each candidate dialogue agent, the semantic similarity between historical text information and other text information can be calculated, and historical text information with a semantic similarity greater than a certain threshold can be regarded as historical text information similar to other text information.

[0142] Understandably, if historical text information similar to the current text exists, it indicates that a user has already performed a search operation on that text information or similar text information before the current time period. The information from that search operation can be used as a reference for the current text information. Therefore, the candidate dialogue agent does not need to generate new response information for the current text information and use it for subsequent evaluation. Instead, it can directly use the historical response information of the historical text information as the response information for the current text information and use it for subsequent evaluation.

[0143] Considering that search results may include information from multiple candidate dialogue agents, and that generating responses from multiple candidate dialogue agents based on text information consumes significant computational resources, this disclosure improves the efficiency of response information acquisition by obtaining historical dialogue information from each candidate dialogue agent. Furthermore, if similar historical text information exists within the historical dialogue information, historical response information can be directly reused. This enhances the evaluation efficiency of candidate dialogue agents and improves the user experience.

[0144] According to another embodiment of this disclosure, the method further includes: when the candidate dialogue agent's historical dialogue information does not include historical text information similar to the text information, invoking the candidate dialogue agent to generate response information for the text information.

[0145] If the historical dialogue information does not include historical text information similar to the current text information, it indicates that no user has performed a search operation for similar text information before the current time period. In order to avoid reducing the evaluation accuracy of the candidate dialogue agent, the candidate dialogue agent can be invoked to generate the corresponding answer information based on the current input text information according to the identification information of the candidate dialogue agent.

[0146] In this embodiment, if there is no historical text information similar to the current text information in the historical dialogue information, the candidate dialogue agent is called to generate response information for the text information, which improves the accuracy of response information in the large model input, and thus improves the evaluation accuracy of subsequent candidate dialogue agents.

[0147] In other embodiments, if the historical dialogue information does not include historical text information similar to the text information, the predetermined response information can be directly obtained and input into the large model. For example, if the predetermined response information is "missing dialogue information," the large model can mainly evaluate candidate dialogue agents based on business information and text information to balance evaluation efficiency and evaluation accuracy.

[0148] According to embodiments of this disclosure, the search results further include search page information for at least one search page, wherein the search page information is determined based on the matching degree between the text information and at least one keyword in the search page information; displaying the trigger entry information of the target dialogue agent further includes: displaying the trigger entry information of the target dialogue agent and the trigger entry information of the search page, wherein the display order of the trigger entry information of at least one target dialogue agent is before the trigger entry information of the search page, and the trigger entry information of the target dialogue agent includes answer information.

[0149] According to embodiments of this disclosure, search page information may be information used to identify the page, such as the retrieval path of the search page.

[0150] Both the search page and the dialogue window of the target dialogue agent mentioned above can serve as the landing page for search results. Unlike the dialogue window of the target dialogue agent, the search page includes various information, such as the name of the entity to which the search results belong, business information, and other information.

[0151] For each search page to be filtered, the search engine can be used to obtain the search page information for each page, which includes at least one keyword from the search page. Then, the search results can be used to determine whether the search page to be filtered is included in the search results based on the matching degree between at least one keyword and the text information. For example, the matching degree between keywords and text information can be semantic similarity. The matching degree between each keyword and text information can be weighted and summed, and the result of the weighted sum can be used to determine whether to return the search page to be filtered as a search page.

[0152] Figure 5 The illustration depicts a scenario where a search page is triggered according to an embodiment of this disclosure. Figure 5As shown, users can input text information through the interactive control 511 in the interactive page 510 of the search engine. In this interactive page 510, search results include the trigger entry information 512 of the target dialogue agent and the trigger entry information 513 of the search page 520. The search page 520 can be the domain of the entity providing the community forum service. The search page 520 includes information such as the article's publication time, publisher, article source, image explanations of keywords in the text information, and comments. The article source can be the name of the domain. All the information displayed on the search page can be used as search page information.

[0153] During the search process, keywords can be extracted from any of the following information: publication time, publisher, article source, image explanation of keywords in the text information, or comments. Based on the matching degree between at least one keyword and the text information, the search results include the current search page 520. Then, the trigger entry information 513 of the search page 520 is displayed on the interactive page 510.

[0154] In this embodiment, the trigger entry information 512 of the target dialogue agent may include the text information "the effect of herb A" and the response information to the text information "a new literature has recently emerged explaining the effect of herb A in the XX field, the new literature is YYY". If there are multiple target dialogue agents, the trigger entry information of multiple target dialogue agents can be displayed before the trigger entry information on the search page.

[0155] In the embodiments of this disclosure, when the search results include both search page information and target dialogue agent information, the trigger entry information of the target dialogue agent can be displayed first to meet the user's needs for immediate answers and interactive communication. Furthermore, displaying the answer information in the trigger entry information allows the user to intuitively and quickly obtain a targeted answer from the target dialogue agent without entering the target dialogue agent's dialogue window. Since this embodiment only requires the user to browse the answer information in the trigger entry information to select the agent, it optimizes the user interaction logic of triggering the target dialogue agent, discovering it is not the desired dialogue agent, and exiting, thereby improving the user experience and information acquisition efficiency.

[0156] In the field of advertising search, the target of candidate dialogue agents can be advertisers. The landing pages of these candidate dialogue agents may not only lack business information, but may also display false information in the trigger entry information, potentially misleading users and affecting their search experience. In the embodiments of this disclosure, by utilizing business information, text information, and response information, not only is the business information required for the candidate dialogue agent evaluation process improved, but the accuracy of the evaluation information is also enhanced, thereby improving the search accuracy of the candidate dialogue agents.

[0157] According to embodiments of this disclosure, the large model can be obtained by fine-tuning a pre-trained large model, such as an LLM, using training data. The training data includes sample input data and sample output data. The sample input data includes sample text information, sample response information, and sample business information. The sample business information can serve as the business information of the candidate dialogue agent to which the sample response information belongs.

[0158] The sample output data includes sample evaluation information, such as sample evaluation values ​​and reasons for evaluation. The sample evaluation information can be obtained based on human scoring, and the reasons for evaluation can also be generated by processing the sample evaluation values ​​and sample input data using other large models.

[0159] Stochastic gradient descent can be used to fine-tune the pre-trained large model using multiple sets of training data, and the large model can be obtained when the loss function converges.

[0160] During fine-tuning, you can first fine-tune the generation task of the pre-trained large model, and then fine-tune the discrimination and generation tasks together. For example, the loss function can include generation loss terms and discrimination loss terms. Alternatively, you can directly fine-tune the discrimination and generation tasks together.

[0161] The generation loss term can be the negative log-likelihood loss between the generated tokens and the input tokens. The input tokens can be tokens obtained by encoding the text in the sample input data. The discriminative loss term can be the cross-entropy loss.

[0162] For example, generating the loss term L gen (U) and discriminant loss term L cls (U) are respectively from formulas (1) and (2).

[0163] (1)

[0164] Where U represents the training data, U includes sample input data and sample output data, U={u1...u n u n+1...u m}, where u1...u n Indicates the sample input data, u n+1 ...u m This represents the sample output data. Indicated by The parameters of the pre-trained large model are based on generate The conditional probability.

[0165] (2)

[0166] in, This represents the parameters of the fully connected layer used for classification in a pre-trained large model. The predicted vector of the word cls located at the last position of the sample input data can be used as input to the fully connected layer for classification. This represents the classification score output after the fully connected layer.

[0167] In one specific embodiment, the loss function L(U) can be expressed as in formula (3).

[0168] (3)

[0169] Formula (3) allows for the simultaneous fine-tuning of both the discrimination and generation tasks.

[0170] In another embodiment, the loss function can be as shown in formula (4).

[0171] (4)

[0172] Where mask represents the moderating variable, This is a hyperparameter used to control the ratio between the generation loss term and the discrimination loss term. The loss function can be dynamically adjusted using the mask. For example, when fine-tuning the generation task of a pre-trained large model first, the mask can be set to 0; when fine-tuning both the discrimination and generation tasks together, the mask can be set to 1. It can be understood that when fine-tuning the generation task, the generation loss term... Previous coefficients It can also be adjusted to 1.

[0173] In the embodiments of this disclosure, by first fine-tuning the generation task of the pre-trained large model, and then fine-tuning the discrimination task and the generation task together, it is possible to improve training efficiency and training effect while improving the generation effect.

[0174] Figure 6 The diagram schematically illustrates a framework for generating evaluation rationale information and evaluation values ​​for a large model according to an embodiment of this disclosure. For example... Figure 6 As shown, the input to a large model can include text information, response information, and business information, etc. After encoding, the input can be encoded into multiple tokens, such as T. ins T0...T 分隔符 ...gmask...T n ,cls, where T ins `cls` can be used to mark the start and end boundaries of the input. The `T` delimiter can be either the first or second delimiter mentioned above, used to identify the relationship between multiple consecutive tokens in the input. `gmask` can be used to indicate whether to fine-tune the generation task first. Multiple tokens can be mapped to obtain corresponding embedding vectors, such as E... ins E0...E 分隔符 ...E gmask ...E n E cls Here, E can refer to the embedding vector. After processing by the large model, the embedding vector can yield multiple predicted output vectors, such as F. ins F0...F 分隔符 ...F gmask ...F n F cls .

[0175] During the fine-tuning phase, `gmask` can be used to adjust whether to fine-tune the generation task first; during the inference phase, `gmask` can be empty. For F... gmask Vector transformation can yield evaluation reasoning information for F. cls The evaluation value can be obtained by performing vector transformation.

[0176] Figure 7 A block diagram of an artificial intelligence-based search device according to an embodiment of the present disclosure is illustrated schematically. Figure 7 As shown, the artificial intelligence-based search device 700 includes:

[0177] The acquisition module 710 is used to acquire search results that match the text information input by the object. The search results include information on multiple candidate dialogue agents. The information on candidate dialogue agents includes the trigger entry information of the candidate dialogue agent, the business information of the object to which the candidate dialogue agent belongs, and the response information of the candidate dialogue agent to the text information generated by the candidate dialogue agent based on the business information.

[0178] The generation module 720 is used to process text information, business information and answer information corresponding to multiple candidate dialogue agents using a large model, and generate evaluation information for each of the multiple candidate dialogue agents.

[0179] The determination module 730 is used to determine the target dialogue agent from multiple candidate dialogue agents based on multiple evaluation information.

[0180] Display module 740 is used to display the trigger entry information of the target dialogue agent.

[0181] According to embodiments of this disclosure, the generation module 720 includes a first filtering submodule, configured to filter at least one first candidate dialogue agent from multiple candidate dialogue agents based on the semantic similarity between text information and business information. The first generation submodule is configured to, for each first candidate dialogue agent, process text information, corresponding answer information, and business information using a large model to generate evaluation information for the first candidate dialogue agent.

[0182] According to embodiments of this disclosure, the evaluation information includes: an evaluation value and evaluation justification information. The first generation submodule includes: a first generation unit, configured to combine the response information corresponding to the first candidate dialogue agent, business information, and text information corresponding to the object into prompt information and input it into a large model, so as to utilize the large model to perform a discrimination task and a generation task respectively, to obtain an evaluation value and evaluation justification information. The evaluation value is used to quantify the degree of matching between the information corresponding to the first candidate dialogue agent and the information corresponding to the object, and the evaluation justification information is used to explain the evaluation value.

[0183] According to an embodiment of the present disclosure, the determining module 730 includes a first determining submodule, used to select a target dialogue agent from a first candidate dialogue agent based on a comparison relationship between an evaluation value and an evaluation threshold.

[0184] According to embodiments of this disclosure, the AI-based search device 700 further includes: a first search module, configured to search a knowledge base for domain background information and / or word explanation information matching the text information, wherein the domain background information is determined based on semantic search of the text information, and the word explanation information is determined based on semantic search of at least one keyword in the text information; a combination module, configured to combine at least one of the domain background information and word explanation information corresponding to the object, and the text information, as information corresponding to the object; and combine the information corresponding to the object with the answer information and business information corresponding to the first candidate dialogue agent into prompt information; and a second generation submodule, configured to input the prompt information into a large model, so as to use the large model to perform a discrimination task and a generation task respectively to obtain an evaluation value and evaluation reason information.

[0185] According to embodiments of this disclosure, the prompt information includes a first separator and at least one second separator. The first separator is used to separate information corresponding to the object and information corresponding to the first candidate dialogue agent. The second separator is used to separate business information and answer information. The second separator is also used to separate at least two of text information, domain background information, and word explanation information. At least one piece of information corresponding to the object and at least one piece of information corresponding to the first candidate dialogue agent in the prompt information are arranged in a predetermined order.

[0186] According to an embodiment of this disclosure, the generation module 720 further includes a simplification submodule, which is used to simplify the answer information based on the business information using the large model before using the large model to process the text information, the answer information corresponding to the first candidate dialogue agent, and the business information to generate the evaluation information of the first candidate dialogue agent, so as to obtain simplified answer information; wherein, the simplified answer information, together with the business information and the answer information, is used to generate the evaluation information.

[0187] According to an embodiment of this disclosure, the simplification submodule includes a first determining unit, configured to determine dynamic text constraints that match the text length of the business information, wherein the dynamic text constraints are used to limit the text length of the response information; and a simplification unit, configured to use a large model to simplify the response information according to the dynamic text constraints to obtain simplified response information.

[0188] According to an embodiment of this disclosure, the generation module 720 further includes a conflict detection submodule, used to perform business conflict detection on the business information using a large model to obtain a business conflict detection result; and a fourth generation submodule, used to process text information, response information corresponding to the first candidate dialogue agent, and at least two business sub-information using a large model when the business conflict detection result indicates that there are at least two conflicting business sub-information in the business information, and generate at least two evaluation information for the first candidate dialogue agent, wherein the evaluation information corresponds to the business sub-information.

[0189] According to an embodiment of this disclosure, the determining module 730 further includes a second determining submodule, which is used to determine the first candidate dialogue agent as the target dialogue agent if, for a first candidate dialogue agent having at least two evaluation information, there is evaluation information in the at least two evaluation information whose evaluation value is greater than the evaluation threshold.

[0190] According to an embodiment of this disclosure, the generation module 720 further includes a fifth generation submodule, used to, when there is a second agent other than the first candidate dialogue agent among multiple candidate dialogue agents, use a predetermined evaluation value and predetermined text information as the evaluation value and evaluation reason information of the second agent, respectively, wherein the predetermined evaluation value is less than or equal to the evaluation value generated using a large model.

[0191] According to embodiments of this disclosure, the candidate dialogue agent information includes the identification information of the candidate dialogue agent. The acquisition module 710 includes: a search submodule, used to acquire the identification information and trigger entry information of each of the multiple candidate dialogue agents that match the text information using a search engine; and a first information acquisition submodule, used to acquire the business information of the home object of each candidate dialogue agent from the home object maintenance library based on the identification information of each candidate dialogue agent.

[0192] According to an embodiment of this disclosure, the acquisition module 710 further includes: a second information acquisition submodule, configured to acquire historical dialogue information of each candidate dialogue agent from the home object maintenance library based on the identification information of each candidate dialogue agent; and a third information acquisition submodule, configured to determine historical text information similar to the text information from the historical dialogue information of each candidate dialogue agent, and to use historical response information to the historical text information as response information to the text information.

[0193] According to an embodiment of this disclosure, the acquisition module 710 further includes: a fourth information acquisition submodule, used to call the candidate dialogue agent to generate response information for the text information when the candidate dialogue agent's historical dialogue information does not include historical text information similar to the text information.

[0194] According to embodiments of this disclosure, the search results further include at least one search page information, wherein the search page information is determined based on the matching degree between the text information and at least one keyword in the search page information; the display module 740 includes: a display submodule, used to display the trigger entry information of the target dialogue agent and the trigger entry information of the search page, wherein the display order of the trigger entry information of the target dialogue agent is located before the trigger entry information of the search page, and the trigger entry information of the target dialogue agent includes answer information.

[0195] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0196] According to an embodiment of the present disclosure, an electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method described above.

[0197] According to embodiments of the present disclosure, a non-transitory computer-readable storage medium stores computer instructions, wherein the computer instructions are used to cause a computer to perform the methods described above.

[0198] According to an embodiment of this disclosure, a computer program product includes a computer program that, when executed by a processor, implements the method described above.

[0199] Figure 8 A schematic block diagram of an artificial intelligence agent according to an embodiment of the present disclosure is shown.

[0200] In embodiments of this disclosure, the von Neumann architecture in modern computer theory is inspired, such as... Figure 8 As shown, the AI ​​agent 800 may include five core modules: input module 810, processing module 820 and output module 830.

[0201] The input module 810 is responsible for receiving or sensing information such as queries, requests, instructions, signals, or data from the outside world (e.g., users or the external environment), and converting it into a format that the AI ​​agent 800 can understand and process. The input module 810 is the primary link for the AI ​​agent 800 to interact with the outside world. It enables the AI ​​agent 800 to efficiently and accurately obtain the necessary "sensory" information from the outside world and respond to this information.

[0202] In the example, the input module 810 can input the text information described above, answer information, business information, word explanation information, domain background information, etc.

[0203] In the example, the processing module 820 is the core support for the AI ​​agent 800's ability to handle complex tasks. The processing module 820 is used to determine the target task based on the input information received by the input module 810, determine the large model based on the target task, and execute the AI-based search method described above by calling the large model.

[0204] In the example, output module 830 can output the evaluation information described above.

[0205] In the example, the processing module 820 may include a control unit 821, a storage unit 822, and a processing unit 823.

[0206] During operation, the control unit 821 will continuously interact with the storage unit 822, the arithmetic unit 823, and / or the output module 830. However, it should be noted that in the embodiments of this disclosure, the control unit 821 initiates communication with the storage unit 822, the arithmetic unit 823, and / or the output module 830 as a single initiator, and there is no communication coupling between the storage unit 822, the arithmetic unit 823, and the output module 830.

[0207] In the example, the performance of the control unit 821 is closely related to the large model on which the AI ​​agent 800 is based. To fully leverage the capabilities of the large model, the internal structure of the control unit 821 can be designed to be highly configurable and scalable to handle various types of tasks and requirements in real-world scenarios.

[0208] Storage unit 822 can be responsible for remembering information such as evaluation values ​​and evaluation reasons.

[0209] The operation unit 823 can be viewed as a predefined tool library. Tools for intra-class space analysis and inter-class space analysis, as described above, can be included in the operation unit 823.

[0210] In the example, after acquiring input information, the AI ​​agent 800 can use the input information to determine the target task and, based on the target task, determine a large model. The input information can be stored in the storage unit 822 of the processing module 820. The control unit 821 can call the large model to retrieve the input information from the storage unit 822 and process it to generate evaluation information, which is then stored in the storage unit 822. And / or the control unit 821 can transmit the evaluation information as output information to the output module 830.

[0211] The AI ​​agent 900 according to embodiments of this disclosure can simply and effectively improve the level of intelligence, as well as enhance flexibility and versatility.

[0212] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0213] According to an embodiment of the present disclosure, an electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method described above.

[0214] According to embodiments of the present disclosure, a non-transitory computer-readable storage medium stores computer instructions, wherein the computer instructions are used to cause a computer to perform the method described above.

[0215] According to an embodiment of this disclosure, a computer program product includes a computer program that, when executed by a processor, implements the method described above.

[0216] Figure 9The diagram schematically illustrates an electronic device suitable for implementing an artificial intelligence-based search method according to embodiments of the present disclosure. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0217] like Figure 9 As shown, the electronic device 900 includes a computing unit 901, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 902 or a computer program loaded from a storage unit 908 into a random access memory (RAM) 903. The RAM 903 may also store various programs and data required for the operation of the electronic device 900. The computing unit 901, ROM 902, and RAM 903 are interconnected via a bus 904. An input / output (I / O) interface 905 is also connected to the bus 904.

[0218] Multiple components in electronic device 900 are connected to input / output (I / O) interface 905, including: input unit 906, such as keyboard, mouse, etc.; output unit 907, such as various types of monitors, speakers, etc.; storage unit 908, such as disk, optical disk, etc.; and communication unit 909, such as network card, modem, wireless transceiver, etc. Communication unit 909 allows electronic device 900 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0219] The computing unit 901 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 901 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 901 performs the various methods and processes described above, such as AI-based search methods. For example, in some embodiments, the AI-based search method can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 908. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 900 via ROM 902 and / or communication unit 909. When the computer program is loaded into RAM 903 and executed by the computing unit 901, one or more steps of the AI-based search method described above can be performed. Alternatively, in other embodiments, the computing unit 901 can be configured to perform AI-based search methods by any other suitable means (e.g., by means of firmware).

[0220] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0221] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0222] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0223] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0224] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0225] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, distributed system servers, or servers incorporating blockchain technology.

[0226] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this disclosure can be achieved, and this is not limited herein.

[0227] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. An artificial intelligence-based search method, comprising: Obtain search results that match the text information input by the object, wherein the search results include information on multiple candidate dialogue agents, the candidate dialogue agent information includes trigger entry information of the candidate dialogue agent, business information of the belonging object, and response information of the candidate dialogue agent to the text information generated by the candidate dialogue agent based on the business information; The text information, business information corresponding to the multiple candidate dialogue agents, and answer information are processed using a large model to generate evaluation information for each of the multiple candidate dialogue agents. Based on the evaluation information, a target dialogue agent is determined from among the candidate dialogue agents; and Display the trigger entry information of the target dialogue agent.

2. The method according to claim 1, wherein, The process of using a large model to process the text information, the business information corresponding to the multiple candidate dialogue agents, and the response information to generate evaluation information for each of the multiple candidate dialogue agents includes: Based on the semantic similarity between the text information and the business information, at least one first candidate dialogue agent is selected from a plurality of candidate dialogue agents; For each of the first candidate dialogue agents, the large model is used to process the text information, the response information corresponding to the first candidate dialogue agent, and the business information to generate evaluation information for the first candidate dialogue agent.

3. The method according to claim 2, wherein, The evaluation information includes: evaluation value and evaluation reasoning information; the step of using the large model to process the text information, the response information corresponding to the first candidate dialogue agent, and the business information to generate the evaluation information of the first candidate dialogue agent includes: The response information corresponding to the first candidate dialogue agent, the business information, and the text information corresponding to the object are combined into a prompt information and input into the large model. The large model is then used to perform a discrimination task and a generation task respectively to obtain the evaluation value and the evaluation reason information. The evaluation value is used to quantify the degree of matching between the information corresponding to the first candidate dialogue agent and the information corresponding to the object, and the evaluation reason information is used to explain the evaluation value.

4. The method according to claim 3, wherein, The step of determining the target dialogue agent from the multiple candidate dialogue agents based on the multiple evaluation information includes: Based on the comparison between the evaluation value and the evaluation threshold, the target dialogue agent is selected from at least one of the first candidate dialogue agents.

5. The method according to claim 2, wherein, The method further includes: Search the knowledge base for domain background information and / or word explanation information that match the text information, wherein the domain background information is determined based on the semantic search of the text information, and the word explanation information is determined based on the semantic search of at least one keyword in the text information; At least one of the domain background information and the word explanation information, along with the text information, shall be used as information corresponding to the object; The information corresponding to the object, the answer information corresponding to the first candidate dialogue agent, and the business information are combined into a prompt information; The prompt information is input into the large model, which is then used to perform the discrimination task and the generation task to obtain the evaluation value and the evaluation reason information.

6. The method according to claim 3 or 4, wherein, The prompt information includes a first separator and at least one second separator. The first separator is used to separate the information corresponding to the object and the information corresponding to the first candidate dialogue agent. The second separator is used to separate the business information and the answer information. The second separator is also used to separate at least two of the text information, the domain background information, and the word explanation information. The information in the prompt message, including at least one piece of information corresponding to the object and at least one piece of information corresponding to the first candidate dialogue agent, are arranged in a predetermined order.

7. The method according to any one of claims 2 to 6, wherein, Before using the large model to process the text information, the response information corresponding to the first candidate dialogue agent, and the business information to generate evaluation information for the first candidate dialogue agent, the method further includes: The large model is used to simplify the answer information based on the business information to obtain simplified answer information, wherein the simplified answer information, together with the business information and the answer information, is used to generate the evaluation information.

8. The method according to claim 7, wherein, The step of simplifying the response information using the large model based on the business information to obtain simplified response information includes: Determine dynamic text constraints that match the text length of the business information, wherein the dynamic text constraints are used to limit the text length of the response information; Using the large model, the answer information is simplified according to the dynamic text constraints to obtain the simplified answer information.

9. The method according to any one of claims 2 to 8, wherein, The step of using the large model to process the text information, the response information corresponding to the first candidate dialogue agent, and the business information to generate evaluation information for the first candidate dialogue agent includes: The large model is used to perform business conflict detection on the business information to obtain business conflict detection results. When the business conflict detection result indicates that there are at least two conflicting business sub-information in the business information, the large model is used to process the text information, the response information corresponding to the first candidate dialogue agent, and the at least two business sub-information to generate at least two evaluation information for the first candidate dialogue agent, wherein the evaluation information corresponds to the business sub-information.

10. The method according to claim 9, wherein, The step of determining the target dialogue agent from the multiple candidate dialogue agents based on the multiple evaluation information includes: For a first candidate dialogue agent that has at least two of the evaluation information, if there is an evaluation information in the at least two of the evaluation information whose evaluation value is greater than the evaluation threshold, the first candidate dialogue agent is determined as the target dialogue agent.

11. The method according to any one of claims 2 to 10, wherein, The method further includes: If a second agent exists among the multiple candidate dialogue agents other than the first candidate dialogue agent, a predetermined evaluation value and predetermined text information are used as the evaluation value and evaluation reason information of the second agent, respectively, wherein the predetermined evaluation value is less than or equal to the evaluation value generated using the large model.

12. The method according to any one of claims 1 to 11, wherein, The candidate dialogue agent information also includes the identification information of the candidate dialogue agent, and the search results that match the text information input by the object include: The identification information and trigger entry information of each of the multiple candidate dialogue agents that match the text information are obtained using a search engine. Based on the identification information of each candidate dialogue agent, the business information of the home object of each candidate dialogue agent is obtained from the home object maintenance library.

13. The method according to any one of claims 1 to 12, wherein, The process of obtaining search results that match the text information input by the object also includes: Based on the identification information of each candidate dialogue agent, the historical dialogue information of each candidate dialogue agent is obtained from the ownership object maintenance library; Historical text information similar to the text information is determined from the historical dialogue information of each candidate dialogue agent, and historical response information to the historical text information is used as response information to the text information.

14. The method according to claim 13, wherein, The method further includes: If the candidate dialogue agent's historical dialogue information does not include historical text information similar to the text information, the candidate dialogue agent is invoked to generate the response information for the text information.

15. The method according to any one of claims 1 to 14, wherein, The search results also include search page information for at least one search page, wherein the search page information is determined based on the degree of matching between the text information and at least one keyword in the search page information; The method of displaying the trigger entry information of the target dialogue agent also includes: The trigger entry information of the target dialogue agent and the trigger entry information of the search page are displayed, wherein the display order of the trigger entry information of the target dialogue agent is before the trigger entry information of the search page, and the trigger entry information of the target dialogue agent includes the answer information.

16. An artificial intelligence-based search device, comprising: The acquisition module is used to acquire search results that match the text information input by the object. The search results include information on multiple candidate dialogue agents. The information on the candidate dialogue agents includes the trigger entry information of the candidate dialogue agents, the business information of the belonging object, and the response information of the candidate dialogue agents to the text information generated by the candidate dialogue agents based on the business information. The generation module is used to process the text information, the business information corresponding to the multiple candidate dialogue agents, and the answer information using a large model, and to generate evaluation information for each of the multiple candidate dialogue agents. The determination module is configured to determine a target dialogue agent from a plurality of candidate dialogue agents based on a plurality of evaluation information; and The display module is used to display the trigger entry information of the target dialogue agent.

17. An electronic device comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1 to 15.

18. An intelligent agent based on artificial intelligence, comprising: The input module is used to receive input information; The processing module is configured to determine a target task based on the input information received by the input module, determine a large model based on the target task, and execute the method of any one of claims 1 to 15 by calling the large model to obtain output information; An output module is used to output the output information obtained by the processing module.

19. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1 to 15.

20. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1 to 15.

Images